text_categorization/cnn_model.py at master · hbahadirsahin/text_categorization · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
import tensorflow as tf


class CNN_Categorizer_v2:
    def __init__(self, sentence_length, num_classes, vocabulary_size, embedding_size, filter_sizes, num_filters,
                 l2_reg_lambda=0.0, embedding_type="static", pretrained_embedding=None):
        self.sentence_length = sentence_length
        self.num_classes = num_classes
        self.vocabulary_size = vocabulary_size
        self.pretrained_embedding = pretrained_embedding
        self.embedding_size = embedding_size
        self.filter_sizes = filter_sizes
        self.num_filters = num_filters
        self.embedding_type = embedding_type

        self.input_x = tf.placeholder(tf.int32, [None, self.sentence_length], name='input_x')
        self.input_y = tf.placeholder(tf.float32, [None, self.num_classes], name="input_y")
        self.keep_prob = tf.placeholder(tf.float32, name='keep_prob')
        self.batch_norm = tf.placeholder(tf.bool, name="batch_norm")

        self.embedding_placeholder = tf.placeholder(tf.float32, [self.vocabulary_size, embedding_size],
                                                    name="pretrained_embedding")

        with tf.device('/cpu:0'), tf.name_scope("embedding"):
            # Training embedding stuff
            if self.embedding_type == "static":
                embedding_ = tf.Variable(tf.constant(0.0, shape=[self.vocabulary_size, embedding_size]),
                                         trainable=False,
                                         name='embedding_weight')
                embedding_train = tf.assign(embedding_, self.embedding_placeholder)
            elif self.embedding_type == "nonstatic":
                embedding_ = tf.Variable(tf.constant(0.0, shape=[self.vocabulary_size, embedding_size]),
                                         trainable=True,
                                         name='embedding_weight')
                embedding_train = tf.assign(embedding_, self.embedding_placeholder)
            elif self.embedding_type == "random":
                embedding_train = tf.Variable(tf.truncated_normal([self.vocabulary_size+1, self.embedding_size],
                                                            name='pretrained_embedding'))
            elif self.embedding_type == "multichannel":
                raise NotImplementedError("No multichannel implementation")
            else:
                print("Invalid empedding type:", self.embedding_type)

            # Validation embedding must be non-trainable!
            embedding_ = tf.Variable(tf.constant(0.0, shape=[self.vocabulary_size, embedding_size]),
                                     trainable=False,
                                     name='embedding_weight_vali')
            embedding_validation = tf.assign(embedding_, self.embedding_placeholder)

            embedding = tf.cond(self.batch_norm, lambda: embedding_train, lambda: embedding_validation)

            self.embedding_input = tf.nn.embedding_lookup(embedding, self.input_x)
            self.static_embedding_input_expanded = tf.expand_dims(self.embedding_input, -1)

        with tf.device('/gpu:0'):
            layers = []
            for i, filter_size in enumerate(filter_sizes):
                with tf.name_scope("conv-maxpool-" + str(filter_size)):
                    conv_filter_shape = [filter_size, self.embedding_size, 1, self.num_filters]
                    pool_filter_shape = [1, self.sentence_length - filter_size + 1, 1, 1]
                    weight = self._get_weights(conv_filter_shape)
                    bias = self._get_bias([self.num_filters])
                    conv = self._set_convolution(self.static_embedding_input_expanded, weight) + bias
                    conv_bn = tf.layers.batch_normalization(conv, training=self.batch_norm)
                    h = tf.nn.relu(conv_bn)
                    pool = self._set_pool(h, pool_filter_shape)
                    layers.append(pool)

            total_num_filters = num_filters * len(filter_sizes)
            self.h_pool = tf.concat(layers, 3)
            self.h_pool_flat = tf.reshape(self.h_pool, [-1, total_num_filters])

            self.h_dropout = tf.nn.dropout(self.h_pool_flat, keep_prob=self.keep_prob)

            with tf.name_scope("fully_connected_layer1"):
                weight_fc1 = self._get_weights(shape=[total_num_filters, 128])
                bias_fc1 = self._get_bias(shape=[128])
                fc1 = tf.nn.xw_plus_b(self.h_dropout, weight_fc1, bias_fc1)
                fc1_bn = tf.layers.batch_normalization(fc1, training=self.batch_norm)
                fc1_out = tf.nn.relu(fc1_bn, name="fc1_out")
                fc1_dropout = tf.nn.dropout(fc1_out, keep_prob=self.keep_prob)

            # with tf.name_scope("fully_connected_layer2"):
            #     weight_fc2 = self._get_weights(shape=[128, 64])
            #     bias_fc2 = self._get_bias(shape=[64])
            #     fc2 = tf.nn.xw_plus_b(fc1_dropout, weight_fc2, bias_fc2)
            #     fc2_bn = tf.layers.batch_normalization(fc2, training=self.batch_norm)
            #     fc2_out = tf.nn.relu(fc2_bn, name="fc2_out")
            #     fc2_drop = tf.nn.dropout(fc2_out, self.keep_prob)

            with tf.name_scope("output"):
                weight_out = self._get_weights(shape=[128, self.num_classes])
                bias_out = self._get_bias(shape=[num_classes])
                output = tf.nn.xw_plus_b(fc1_dropout, weights=weight_out, biases=bias_out, name="logits")

            with tf.name_scope("loss"):
                losses = tf.nn.softmax_cross_entropy_with_logits_v2(labels=self.input_y, logits=output)
                l2_losses = tf.add_n([tf.nn.l2_loss(tf.cast(v, tf.float32)) for v in tf.trainable_variables()],
                                     name="l2_losses") * l2_reg_lambda
                losses = tf.reduce_mean(losses, name="softmax_loss")
                self.loss = tf.add(losses, l2_losses, name="loss")

            # Accuracy
            with tf.name_scope("accuracy"):
                self.predictions = tf.argmax(output, 1, name="predictions")
                correct_predictions = tf.equal(self.predictions, tf.argmax(self.input_y, 1))
                self.accuracy = tf.reduce_mean(tf.cast(correct_predictions, tf.float32), name="accuracy")

            with tf.name_scope("accuracytopk"):
                self.pred_top_2 = tf.reduce_mean(
                    tf.cast(tf.nn.in_top_k(self.predictions, tf.argmax(self.input_y, 1), k=2), tf.float32), name="top2")
                self.pred_top_3 = tf.reduce_mean(
                    tf.cast(tf.nn.in_top_k(self.predictions, tf.argmax(self.input_y, 1), k=3), tf.float32), name="top3")
                self.pred_top_4 = tf.reduce_mean(
                    tf.cast(tf.nn.in_top_k(self.predictions, tf.argmax(self.input_y, 1), k=4), tf.float32), name="top4")
                self.pred_top_5 = tf.reduce_mean(
                    tf.cast(tf.nn.in_top_k(self.predictions, tf.argmax(self.input_y, 1), k=5), tf.float32), name="top5")

    def _set_convolution(self, input, weight, strides=[1, 1, 1, 1], padding="VALID"):
        return tf.nn.conv2d(input, weight, strides=strides, padding=padding, name="conv")

    def _set_pool(self, input, shape, strides=[1, 1, 1, 1], padding='VALID'):
        return tf.nn.max_pool(input, ksize=shape, strides=strides, padding=padding, name="pool")

    def _get_weights(self, shape):
        return tf.Variable(tf.truncated_normal(shape, stddev=0.1), name="w")

    def _get_bias(self, shape):
        return tf.Variable(tf.constant(0.1, shape=shape), name="b")